Guideline Representation Formalism and Electronic Decision Support Systems: Addressing the Guideline – Implementation Gap - Part 1

• Abstract
• Introduction
• Guideline Representation and Comprehension Mismatch
• Guideline Representation Formalisms
  • Model-Centric Representation
  • Document-Centric Representation
• Model-Centric Guideline Representation
  • Primitive-based Modelling
  • Problem-solving Method (PSM)-Based Modelling
• Document-Centric Guideline Representation
• Electronic Guideline Representation: Requirements and Issues

Click here for Part 2 of this paper

Abstract
The health care industry and professional colleges around the world have been working laboriously to develop and implement evidence-based clinical guidelines as a mechanism to reduce variability in clinical practice and to improve the quality of health care. For clinical guidelines to impact positively on clinical practices and outcomes, they must be integrated into clinical workflow and decision support applications. Electronic guideline representation formalisms can be divided into document-centric and (execution) model-centric. However, significant variability in guideline representation models and inexact mapping from experts’ internal knowledge to the external representations as narratives and then to machine executable formats have reduced the usability of clinical guidelines both at human and machine implementation levels. This paper analyses different clinical guideline representation formalisms and proposes a tri-model guideline representation architecture and scenario-based authoring methodology to address the problems.

Introduction
The health care industry has been under increasing pressures to reduce health care practice variability and improve quality of services. The two reports by the Institute of Medicine – "To Err is Human: Building a Safer Health System" ^{[ 1 ]} and "Crossing the Quality Chasm" ^{[ 2 ]} have escalated such pressures. These drivers have significantly pushed health care professionals to adopt evidence-based practices. Facing the almost exponential growth in biomedical knowledge, practitioners today have huge difficulty in keeping up with the enormous amount of new information continually becoming available and incorporating the information into their own knowledge to support the making of accurate diagnostic and high quality clinical management decisions. The distillation of the huge amount of internal and external knowledge into evidence-based clinical guidelines is a logical solution to help overcome this serious knowledge–human cognition mismatch.^{[ 3,4 ]}

However, the adoption of clinical practice guidelines in clinical care has traditionally been relatively low.^{[ 5-8 ]} The failure to integrate guidelines into workflow, and their perceived irrelevance to the actual practice of medicine, were some of the reasons for resistance to guideline adoption by clinicians in clinical practice. The development of evidence-based guidelines is both intellectually resource intensive, making guideline development an expensive activity. Rapidly changing evidence and biomedical knowledge add another dimension of difficulty – how to address the need for speedy updates and distribution of the changes in knowledge to clinicians.

The informatics community has been working laboriously on electronic tools to facilitate speedy guideline authoring, editing/revision and distribution. However, in order for clinical practice guidelines to impact positively on clinical care, they need to be integrated seamlessly into the clinical workflow. Guideline-based decision support applications integrated into clinical information systems have been known to improve the quality of decision making in many industries including health care.

A number of international projects have been established to develop computerised guideline implementation and decision support systems.^{[ 9-10 ]} Each has adopted a different approach in guideline representation architecture and implementation. Although the diversity reflects the responsible groups’ different interests and expertises and stimulates creativity, the almost complete absence of standards has created major difficulties for guideline implementers and decision-support systems designers.

The international health informatics community realises the benefit of standards and has expressed interest in developing a standardised approach to guideline representation and implementation at a number of recent Health Level Seven (HL7)^{[ a ]}technical workgroup meetings. However, some difficulties exist, at the technical and political levels in particular.

The aim of this paper is to explore some of these issues and discuss ways of moving towards standardised architectures for computer interpretable clinical guideline representation.

Guideline Representation and Comprehension Mismatch
Clinical guidelines are developed by experts in a clinical domain using processes of reviewing evidence-based knowledge in the literature, meta-analysis, decision analytic modelling (internal representation) and expert consensus. This internal knowledge is then represented externally using symbols such as written texts, which may be accompanied by algorithmic flowcharts. A person’s prior knowledge and experiences often affect the relationship between internal and external representation. Experts approach guidelines with a more highly organised knowledge base and experience, which may result in guidelines being written at higher degrees of "abstraction". Guidelines are generally developed to provide "population-based" recommendations. As such, they tend to be relatively generic in nature. Many inferences (such as inferencing a disease from a cluster of clinical findings), assumptions and presumptions can also be left out of the external representations.^{[ 11 ]}

As an individual tries to comprehend the text-based clinical guidelines, he/she attempts to infer the knowledge in the text message to construct a cognitive model. This process is known as "knowledge-based inferences". Clinical guidelines can be semantically complex and may contain collections of knowledge abstractions with logical gaps that can promote ambiguity. The inference processes are used "fill the gaps of missing information in the text and sometimes to reorganise the information". This creates a situation known as guideline representation and comprehension mismatch.

Prior knowledge and experiences have been known to significantly influence the way clinicians contextualise external representations (written texts)^[3,11,12] and hence are critical contributors to variations in guideline comprehension by clinicians. Clinical guidelines are used by clinicians with a wide range of knowledge and experiences. Guideline representation and comprehension mismatch significantly compromises the ability of clinical guidelines to decrease variability in clinical care/practice.

Based on the extensive research by Patel et al since the 1980s, it appears that the free text clinical practice guidelines/protocols representation and the generic nature of guidelines fail to provide clinicians with an unambiguous knowledge structure through which to facilitate accurate evidence-based clinical decisions. Erroneous inferences by clinicians often result from interpretation of ambiguous representations in text guidelines, which can be further compromised by inadequate prior knowledge in the clinical domain. Transcription of knowledge from free text clinical guideline into computerised knowledge bases and inference rules to support decision support system design often further perpetuates the interpretation/inference problems especially when the transcription is performed by technical people,^{[ 11 ]} thus affecting the accuracy and usefulness of such systems.

Guideline Representation Formalisms
For clinical guidelines to change practice behaviours and improve quality of care, knowledge embedded in the guidelines must be easily accessible to all who need it in a timely manner at the point of care. As clinical knowledge changes quickly with new scientific discoveries, guideline contents must be updateable and must be distributed easily and speedily. Ideally, they should be able to deal with exceptions (ie, patient specific conditions) and be integrated seamlessly into the clinical workflow. ^{[ 12 ]} The need for rapid access to guideline knowledge contents and to minimise (or eliminate) “guideline representation–comprehension mismatch” have driven researchers to exploit information technology to develop electronic guideline-based decision support applications that are integrated to clinical information systems in order to overcome these pitfalls of paper-based clinical guidelines.^{[ 13 ]} Many electronic guidelines as well as decision support systems incorporating different guideline architectures have been developed.^{[ 14 ]}

Electronic guideline representation formalisms can generally be classified into two main categories: model-centric representation and document-centric representation.

Model-Centric Representation
This involves the gradual conversion of a conventional narrative guideline formulated by domain experts into a compact conceptual model that is machine interpretable.^{[ 15 ]} However, the "text-to-model" relationship is indirect and the mapping is often inexact. Nevertheless, the conceptual model is semantically close to the operational model; it is much easier for the guideline knowledge components to be used in guideline-based decision support applications.

Document-Centric Representation
A document presentation model is first created. The narratives within the clinical guidelines are then systematically marked-up with XML (eXtensible Markup Language) tags according to the document structure defined by the model. The end product is usually some form of "structured document". Depending on the steps and the transformation (from narrative to XML mark-up) involved in each step, the risk of information loss can vary. Obviously, the lesser the transformation -induced information loss, the easier the subsequent verification.^{[ 15 ]}

Model-Centric Guideline Representation
The (execution) model-centric approach has been a popular formalism for guideline-based decision support system researchers and developers. Most model-centric projects adopt the format of hierarchical decomposition of guidelines into networks of component tasks that unfold over time. This representation has been described as the "Task-Network Model".^{[ 16 ]} The key goal of this approach is to have the guideline knowledge represented as a network of executable tasks. Examples include Arden Syntax, PROforma, GLIF, EON and Prodigy.

This formalism is dominated by two approaches – the primitive-based approach and the problem-solving methods (PSM)-based approach. GLIF, EON and Prodigy all use the knowledge engineering environment of Protégé to encode knowledge components of clinical guidelines. PROforma has a commercial guideline implementation engine and toolset – Arezzo. Most model-centric projects (with the exception of Arden Syntax and Prodigy) are generally experimental projects.

Primitive-based Modelling
Primitives are "concepts" or "classes" for guideline entities (such as "eligibility (or decision) criteria", "actions", "decisions" and "plans") and their attributes. They represent stereotypical tasks defined in a guideline. In primitive-based modelling, guidelines are often represented at a single level of details. Domain and procedural knowledge are often intertwined^{[ 17 ]} making reuse of knowledge slightly more difficult.

Arden Syntax^{[ 18 ]}, PROforma^{[ 16,19 ]} and GLIF^{[ 4 ]} are examples of primitive representation models with syntax for encoding guideline knowledge, decisions, branching and actions. Asbru^{[ 9 ]} , Asgaard^{[ 20 ]} and EON^{[ 21 ]} also provide languages for reasoning with complex temporal logic. In PROforma guidelines are modelled as constraint-satisfaction graphs, where nodes represent tasks (such as clinical actions, decisions or plans), and arcs ("lines")connect tasks within plans. Asbru models guidelines as time-oriented, intention-based plans that can be hierarchically decomposed into (sub)plans or actions. EON enables specification through a combination of modelling primitives such as different types of decision-making mechanisms, control-flow constructs, actions, activities and abstractions.

Problem-solving Method (PSM)-Based Modelling
This formalism considers that decision support systems in general can be split into two independent classes of reusable components: the domain ontologies and problem-solving methods (PSM). Domain ontologies model domain-specific knowledge in terms of concepts/entities, attributes and concept relations.^{[ 22 ]} PSM represents generic strategies to solve stereotypical tasks independent of the system’s application domain.^{[ 17 ]} Well-known PSM-based guideline representation examples include CommanKADS,^{[ 23 ]} OCML,^{[ 24 ]} Protégé^{[ 21 ]} and UOML.^{[ 25 ]}

The PSM-based approach may provide knowledge acquisition tools to facilitate authoring of guidelines by clinical domain experts. Most guideline representation does not use the PSM approach as PSM knowledge acquisition tools are considered incapable of meeting clinical domain experts’ needs and the protocols used in clinical practice do not easily fit PSM’s highly structured formats.^{[ 26 ]} They are more easily expressed in the form of primitive actions.

Document-Centric Guideline Representation
GEM (Guideline Element Model) is a guideline document model.^{[ 5 ]} Guideline knowledge elements are organised hierarchically and marked up with XML tags. The element tags demarcate texts and provide labels that characterise the semantic content of the elements. A GEM specific editor (GEM Cutter) has been developed for marking up text-based guidelines according to the GEM document type definitions (DTD) in the earlier version of GEM. GEM II replaces DTD with an XML schema. An XML schema uses XML syntax and supports more robust data type definitions, making it possible to describe accurately the permissible document contents and to validate the correctness of the data.

The Clinical Practice Guideline Architecture (CPGA) is developed to provide a structure for organising guideline elements in a hierarchy (using XML) to facilitate easy encoding, revising, sharing and viewing of the knowledge contents within the guidelines. The project aims at providing architecture and tools for different types of authors (knowledge and technical) to encode the guideline elements.

Both GEM and CPGA encoded elements can then be grouped, linked to the evidence and source literature, viewed and published. The guideline contents are designed to be computable.

At the September 2004 HL7 Technical Workgroup meeting, the UK University of Newcastle project team reported that CPGA had been replaced by Clinical Practice Guideline Reference Architecture (CPG-RA). CPG-RA is an evidence-based guideline structure work program established within Guideline International Network (GIN). The project team has documented a set of use cases and reviewed the current practices in organising evidence. The use cases are designed for iteratively validating CPG-RA and the guideline knowledge. The clinical guideline structure is then captured using an XML schema.

CPG-RA is designed to formalise the publication of the guideline within a hierarchical, semi-formal model, presented in Microsoft VISIO. Current activities in CPG-RA focus on (a) preparing XML markups and evidence tables to create formalism for organisations which publish guidelines, (b) presenting top level use case for guideline publication, implementation and management and (c) developing "order entry" sets schema.

Electronic Guideline Representation: Requirements and Issues
For electronic guideline representation formalism to be truly robust and useful, the representation must satisfy a number of key criteria:

• Provision of a robust, consistent and technologically independent/neutral structure for modelling narrative guidelines as "structured documents" such that:
• Knowledge elements (eg, the "action", "decision", ["branch steps", "synchronisation steps"] and "plan") within these documents, and their relationships among these elements, are clearly identified
• The knowledge elements can be easily reused in different guidelines
• The structured guideline documents can be distributed to, shared and reviewed by international peers.
• The guideline elements should be:
• Appropriately grouped/organised (eg, in a hierarchy) such that the knowledge elements can be easy and rapid viewed based on patient/population characteristics, eg, age, gender, ethnicity, morbidity, etc
• Linked to evidence sources/literature
• Properly encoded in well-structured formats with computable criteria for execution in/used by decision support applications.
• The guideline representation structure should be appropriate for all types / classes of guidelines, eg, diagnostics, treatment, procedures, etc.
• The guideline structure should support the development, updating of guidelines and encoding of guideline elements by multiple authors and by authors with varying levels of skill.

Extensive literature search efforts lead to the identification of more than 10 guideline representation formalisms and their related applications. Classification and a brief analysis of these formalisms were given in the previous sections. Critiques by other international colleagues can be found at: http://www.openclinical.org/gmmintro.html. A major issue with most of these representation formalisms and applications is that they are not implementation or technologically neutral (independent).

Arden Syntax is a rule-based formalism that represents guidelines as executable "if–then" logic. Medical logic modules of Arden Syntax contain production rules that relate input conditions (eg, patient clinical data) to a particular set of recommended actions. There is no further qualifying clinical domain knowledge in the rules’ premises; hence, reasoning on the clinical concepts and the strategies used to solve the domain problems can be difficult. ^{14 ]} It is incomplete as a structure for representing, especially multi-step practice guidelines that unfold over time. Therefore, it is considered inadequate for modelling care pathways and clinical workflow.^{[ 16 ]}

GLIF is intended as a general architecture for representing computable and sharable guideline elements. There is no provision for narrative guidelines because the emphasis of GLIF is on the detailed specification of guideline recommendations in a structured format. It has not been accepted as generalised guideline representation architecture as it is considered to be non-implementation/technologically independent. The GLIF team has shifted its focus to creating a versatile modelling language that will allow guideline models to be shared between different institutions and software platforms.^{[ 9 ]}

GEM was intended as a document-structuring model for representing narrative clinical guidelines and for rendering guideline elements in the XML documents into structures usable for computer execution. Agrawal & Shiffman^{[ 27 ]} had demonstrated the successful mapping of GEM encoded guideline elements (from the American Academy of Paediatrics guideline on the diagnosis, treatment and evaluation of urinary tract infection in febrile infants and young children) to the Arden Syntax medical logic module.

GEM has been criticised for using "non-specific syntax external to the XML standard" and containing GLIF elements in the GEM DTD (email comm, Ian Purves from the University of Newcastle, ian.purves@ncl.ac.uk, 27 January 2003). The use of non-XML standard syntax and the inclusion of GLIF elements (which are implementation-specific) in the GEM DTD render GEM encoded elements unusable by other systems without human input to transformation. Concern was also raised about the GEM logic components and its requirement to be viewed along with condition elements to determine the sequence and relationship of actions and whether this captured the information as well as a schema-based approach.

Little has been published about CPGA. There is insufficient independent publication on this work to allow an objective evaluation of this guideline representation architecture. The CPGA project team has released a set of documents on its website, see http://www.schin.ncl.ac.uk/cpga/. The documents (which the project team labelled "Use Cases") listed the functional requirements for clinical guideline representation architecture and CPGA XML schema. The "Use Cases" are similar to the computer guideline model key requirement criteria listed at the beginning of this section. An initial examination of the XML schema reveals that the proposed CPGA architecture is non-linear in structure (not unlike that of GEM). It is not clear how this structure would handle branching decision steps.

It appears that CPGA is in the process of being reengineered into CPG-RA. This could present a golden opportunity for the harmonisation of GEM and CPG-RA into a single standard guideline representation architecture. However, disagreements over some technical issues are likely to foil any attempts to achieve this goal.

Footnote
Health Level Seven (HL7) is one of several American National Standards Institute-accredited standards developing organisations (SDOs) operating in the healthcare arena. Most SDOs produce standards (sometimes called specifications or protocols) for a particular healthcare domain such as pharmacy, medical devices, imaging or insurance (claims processing) transactions. HL7’s domain is clinical and administrative data. The HL7 mission is: "To provide standards for the exchange, management and integration of data that support clinical patient care and the management, delivery and evaluation of healthcare services. Specifically, to create flexible, cost effective approaches, standards, guidelines, methodologies, and related services for interoperability between healthcare information systems."

Click here for Part 2 of this paper